Tighten the assistant-tail repair after review so only the Anthropic provider is treated as prefill-rejecting, and so assistant-tail model metadata is evaluated independently from the last user model.
Add regression coverage for an allowed user model followed by a rejecting Anthropic assistant tail, plus a non-Anthropic provider carrying a Claude-looking model id.
Plan: plans/fix-anthropic-assistant-prefill-tail.md
Claude Opus 4.7 rejects requests whose final message is an assistant turn, because Anthropic treats that as unsupported assistant prefill. Add a transform guard that appends the existing synthetic user recovery turn only for Anthropic prefill-rejecting model IDs, while preserving assistant-prefill behavior for supported and unknown models.
Tests cover Opus 4.7, Opus 4.6 assistant metadata fallback, direct Sonnet 4.6 model fields, allowed OpenAI/Sonnet 4.5 tails, missing metadata, and compaction continuation compatibility.
Plan: plans/fix-anthropic-assistant-prefill-tail.md
Discord support and other guidance still tell users 'bunx oh-my-opencode setup', but the CLI only registers 'install'. Running the recommended command produces:
error: unknown command 'setup'
which dead-ends users who are trying to refresh missing skills (hyperplan, etc.) after upgrading. Commander.js supports first-class aliases, so declaring '.alias("setup")' on the install command makes both names route to the same action without any code duplication. Help output now reads 'Usage: oh-my-opencode install|setup [options]', which is the right discovery hint for both newcomers and users following older docs.
Regression test parses cli-program.ts and asserts the install command block contains '.alias("setup")' before its '.action(' so a future refactor of the CLI cannot silently drop the alias.
ensureBaseDirs unconditionally called chmod(baseDir, 0o700) on every startup
and on every team_create. On filesystems where the OS rejects chmod for the
directory (network mounts, SIP-protected locations, non-owner cases on macOS
shared by multiple GUI users), the call raises EPERM and the entire team-mode
init aborts:
[team-mode] init failed: EPERM: operation not permitted, chmod '/Users/<u>/.omo'
Wrap chmod through a small safeChmod helper that converts EPERM, ENOTSUP, and
EINVAL into a single warning log and continues. mkdir already creates new
directories with mode 0o700, and the existing post-creation stat-guard remains
in place for the case where the directory pre-exists with a different mode and
chmod is permitted, so the security envelope on supported filesystems is
unchanged. All other error codes (ENOENT, EACCES, etc.) still propagate.
Regression test mocks node:fs/promises.chmod to throw EPERM and asserts that
ensureBaseDirs completes successfully and emits exactly the documented warning.
The multimodal-looker prompt described what to do but never told the model which tools are available. Smaller VL models (e.g. Qwen3-VL-8B) would try to call non-existent tools and enter an infinite loop emitting:
Model tried to call unavailable tool 'invalid'. Available tools: call_omo_agent, read.
Add a single sentence at the top of the prompt that explicitly enumerates the only allowed tools ('read' and 'call_omo_agent') and forbids calling any other tool. This matches the runtime allowlist enforced by createAgentToolAllowlist(["read"]).
Regression test asserts the prompt contains the available-tools enumeration so future prompt rewrites don't regress.
Require the latest assistant tool-wait turn itself to be stale before a shouldReply parent wake can bypass tool-call deferral. This prevents an all-complete background wake from forking a second parent prompt loop when OpenCode has repaired the tail to a synthetic user message.
Tests:
- bun test src/features/background-agent/parent-wake-user-message-race.test.ts src/features/background-agent/task-completion-cleanup.test.ts src/hooks/shared/prompt-async-gate.test.ts src/shared/prompt-async-route-audit.test.ts --bail
- bun run typecheck
- bun test
Keep the built-in lsp MCP registered even when the submodule CLI artifact is missing.
The fallback command initializes the lsp-tools-mcp submodule, prefers the source CLI without dirtying the checkout with dist output, and keeps npm build as a last resort when Bun cannot run the source entrypoint.
Plan: plans/fix-lsp-mcp-missing-cli.md
The bundled createTodoDescriptionOverrideHook returns { 'tool.definition': fn }, but plugin-interface.ts never exposes 'tool.definition' as an OpenCode hook handler. Result: the hook is constructed by createToolGuardHooks (line 132-134 of src/plugin/hooks/create-tool-guard-hooks.ts) but the function is never invoked, so todowrite keeps using OpenCode's core 7 KB description instead of the 1.4 KB TODOWRITE_DESCRIPTION. User-defined plugins under ~/.config/opencode/plugin/*.js use the same hook contract and work fine, confirming the contract itself is functional in opencode 1.14.28+.
Fix: add src/plugin/tool-definition.ts (createToolDefinitionHandler) that forwards the OpenCode 'tool.definition' input/output pair into hooks.todoDescriptionOverride. Wire it into plugin-interface.ts alongside tool.execute.before/after.
Regression coverage: src/plugin/tool-definition.test.ts covers (a) todowrite override applied, (b) other tools left untouched, (c) null hook is a no-op.
After the 4.2.0 unified-dispatch refactor (a42f894f / df198d8b / fee515c5 / 989ab717 / dd3fecaf / 1bbe065c / 12bd6580), at least one caller in the new prompt-async-gate path forwards a FallbackModelObject (or some other non-string shape) into parsers that statically claim 'model: string'. The downstream .trim() call then throws 'model.trim is not a function', which rejects the session.processor promise and surfaces as 'Aborted process' + UI 'interrupted'. The issue (#4145) reports this aborts 90% of subagent dispatches across every provider on 4.2.0 + opencode 1.15.4.
This patch adds a 'typeof x !== "string"' runtime guard at the four parser entrypoints called from the dispatch path:
- src/shared/fallback-chain-from-models.ts :: parseVariantFromModel, parseFallbackModelEntry
- src/tools/delegate-task/model-string-parser.ts :: parseVariantFromModelID, parseModelString
- src/shared/model-string-parser.ts (duplicate file with same API) :: parseVariantFromModelID, parseModelString
- src/features/claude-code-agent-loader/claude-model-mapper.ts :: mapClaudeModelString
Each parser now returns undefined / { modelID: "" } for non-string input instead of throwing. This unblocks subagent dispatch and leaves the underlying caller bug for a follow-up.
Regression coverage: three new tests in src/shared/fallback-chain-from-models.test.ts pin the non-string behavior (object, null/undefined, number). Existing 38 tests still pass. Total: 41/41 green, typecheck clean.
Skills with an `agent` frontmatter field are intended for a specific
agent. Previously they still appeared in:
- every agent's system prompt (via `buildAvailableSkills`)
- the `skill` tool's `<available_items>` description visible to all agents
This wasted tokens and could mislead agents into attempting calls that
would be rejected at execution time.
Changes:
- `buildAvailableSkills`: new optional `agentName` parameter; when
provided, skills whose `definition.agent` does not match are excluded
- `builtin-agents.ts`: pass per-agent name to `buildAvailableSkills`
for sisyphus, hephaestus, and atlas, so each agent's prompt only
lists the skills it is allowed to use
- `createSkillTool` (`tools.ts`): exclude agent-restricted skills from
both the eager and lazy description builds, keeping the shared tool
description free of skills the current agent cannot access
Execution-time enforcement (throwing on mismatch) is unchanged; this
change adds the earlier, description-level visibility gate.
Tests: new `available-skills.test.ts` (5 cases) + 3 new cases in
`tools.test.ts` covering the description-filter and execute paths.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
When OpenCode reports the original assistant error after a fallback retry has already been accepted, keep waiting for the pending fallback model instead of clearing the awaiting flag.
This prevents a duplicate stale session.error from advancing the fallback chain and dispatching a second assistant retry prompt.
Oracle review on PR #4132 flagged that resetStallStateForPlanChange only
reset state when stalledContinuationReason was already set, so partial
in-progress counts from plan A could carry to plan B and abort the new
plan on its first idle.
Adds activeContinuationPlanPath to SessionState and resets the
no-tool-progress counter + awaiting state whenever the active plan path
changes, even when no stall has occurred yet. The existing stalled-plan
reset still fires when the previously stalled plan path differs from
the new one.
Background fallback retry notifications were queued as bare internal user messages, so OpenCode could treat the notification as a new default-agent turn. Reuse the same parent prompt context resolver used by completion notifications for retrying and retry-ready wakes, and pin regression coverage for Hephaestus parent sessions plus missing-context fallbacks.
buildShellAwareGitPrefix incorrectly returned raw bash VAR=value prefix for csh, which does not support inline env assignment. Now routes csh through buildEnvPrefix() to emit setenv syntax. Also sets code block lang to csh and skips bash block regex prefixing for csh.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
YeonGyu-Kim